Conference on Neural Information Processing Systems · 1874 papers
Multi-turn Reinforcement Learning with Preference Human Feedback
Lior Shani (Google), Remi Munos
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A multi-turn reinforcement learning from preference feedback (MTPO) algorithm is proposed to address the limitations of traditional single-turn RLHF in multi-turn dialogues.
Multi-view Masked Contrastive Representation Learning for Endoscopic Video Analysis
Kai Hu (Xiangtan University), Xieping Gao (Hunan Normal University)
CodeClassificationObject DetectionSegmentationRepresentation LearningTransformerContrastive LearningVideoBiomedical Data
🎯 What it does: A multi-view occlusion contrastive learning framework (MCRL) is proposed for unsupervised pre-training of endoscopic videos, combining global view aggregation attention-guided pipeline occlusion and local view random pipeline occlusion, integrating the occlusion reconstruction task with contrastive learning to enhance pixel-level and global discriminative capabilities.
Multidimensional Fractional Programming for Normalized Cuts
Yannan Chen (Chinese University of Hong Kong), Kaiming Shen (Chinese University of Hong Kong)
CodeSegmentationOptimizationImage
🎯 What it does: A new NCut (Normalized Cut) clustering algorithm (FPC) is proposed through Multidimensional Fractional Programming. This algorithm simplifies the original 0-1 ratio optimization problem into an iteratively solvable linear search problem through a matrix form quadratic transformation.
Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization
James Oldfield (Queen Mary University of London), Ioannis Patras (Queen Mary University of London)
CodeMixture of ExpertsVision Language ModelImageText
🎯 What it does: This paper proposes a multi-linear expert mixture layer µ MoE, utilizing tensor decomposition to achieve scalable expert specialization, addressing the training instability and parameter efficiency issues of sparse MoE.
CodeGenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper generates a large-scale fine-grained human preference dataset, VisionPrefer, using a multimodal large language model (GPT-4 V), and based on this, trains a reward model, VP-Score, for RLHF fine-tuning of text-to-image generation models, thereby enhancing the adherence to prompts, aesthetics, authenticity, and safety of the generated images.
Brandon Huang (University of California), Roei Herzig (IBM Research)
CodeClassificationRecognitionTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: This paper proposes the Multi-modal Task Vector (MTV), which achieves many-shot context learning in a multi-modal setting without being limited by context length by embedding the average activation of multiple examples in the attention heads of large multi-modal models.
MultiOOD: Scaling Out-of-Distribution Detection for Multiple Modalities
Hao Dong (ETH Zurich), Olga Fink (EPFL)
CodeAnomaly DetectionVideoMultimodalityBenchmark
🎯 What it does: This paper proposes a Multi-modal Out-of-Distribution detection (MultiOOD) framework and establishes the first multi-modal OOD benchmark.
Multivariate Probabilistic Time Series Forecasting with Correlated Errors
Vincent Zhihao Zheng (McGill University), Lijun Sun (McGill University)
CodeRecurrent Neural NetworkTransformerTime Series
🎯 What it does: A pluggable method is proposed to learn and calibrate the cross-step error covariance in multivariate probabilistic time series forecasting, thereby improving prediction and uncertainty quantification.
Multivariate Stochastic Dominance via Optimal Transport and Applications to Models Benchmarking
Gabriel Rioux (Center for Applied Mathematics Cornell University), Youssef Mroueh (IBM Research)
CodeRecommendation SystemOptimizationComputational EfficiencyLarge Language ModelTabularBenchmark
🎯 What it does: A multivariate first-order stochastic dominance (FSD) testing method based on entropy regularization is proposed, along with its statistical inference (CLT, bootstrap) and algorithm implementation. Experiments are then conducted on synthetic data and real LLM multi-metric evaluation data.
MutaPLM: Protein Language Modeling for Mutation Explanation and Engineering
YIZHEN LUO, Zaiqing Nie (Pharmolix Inc.)
CodeRecommendation SystemOptimizationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataChain-of-Thought
🎯 What it does: This paper presents MutaPLM, a unified framework based on protein language models for providing interpretable explanations of protein mutations and engineering-recommendable variants.
Abhineet Agarwal (University of California Berkeley), Justin Whitehouse (Carnegie Mellon University)
CodeReinforcement LearningGraph
🎯 What it does: This paper proposes two exploration-commitment algorithms based on a sparse network interference model and discrete Fourier analysis for the multi-armed bandit (MAB) problem in the presence of network interference, aiming to achieve low regret.
🎯 What it does: A mutual information estimation framework based on f-divergence variational representation (f-DIME) is proposed, introducing a derangement sampling strategy to effectively generate marginal distribution samples, significantly reducing estimation variance and improving estimation accuracy.
🎯 What it does: A unified 3D generation framework called MVGamba is proposed, which combines a multi-view diffusion model with an efficient Gaussian reconstructor based on a state space model (SSM) to generate high-quality 3D content in a single forward pass.
Caroline Wang (University of Texas at Austin), Peter Stone (University of Texas at Austin)
CodeReinforcement LearningAgentic AI
🎯 What it does: This paper proposes the N-agent spontaneous team (NAHT) problem, studying how multiple agents can cooperate in the presence of unknown team members.
🎯 What it does: This paper proposes the ChemFlow framework, which views the latent space of molecular generation models as a dynamic system of flow, utilizing vector fields to migrate molecular distributions in the latent space, thereby achieving property optimization and structural exploration.
🎯 What it does: The study applies Dynamic Sparse Training (DST) to extreme multi-label classification (XMC) by using a semi-structured fixed fan-in sparse matrix in the classification layer and introducing auxiliary objectives (meta-classifier or intermediate layer) to improve gradient flow, resulting in significant memory savings during end-to-end training.
🎯 What it does: This paper studies the limitations of parametric dimensionality reduction methods in preserving local structure and proposes a new method called ParamRepulsor to address this issue.
Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models
ShengYun Peng (Georgia Tech), Duen Horng Chau (Georgia Tech)
CodeSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies the security risks of large language models (LLMs) during the fine-tuning process and proposes assessing security by visualizing the safety landscape of model parameter space.
Nearest Neighbor Speculative Decoding for LLM Generation and Attribution
Minghan Li (Cohere), Xi Victoria Lin (Meta)
CodeGenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: NEST is proposed, a semi-parametric language model that performs nearest neighbor speculative decoding during inference to enhance the factuality and source attribution of large language models.
Nearly Optimal Approximation of Matrix Functions by the Lanczos Method
Noah Amsel (New York University), Christopher Musco (University of Massachusetts Amherst)
CodeOptimization
🎯 What it does: This paper studies the application of the Lanczos method in matrix function approximation, particularly proving the approximate optimality of Lanczos-FA when dealing with a class of rational functions.
Nearly Tight Black-Box Auditing of Differentially Private Machine Learning
Meenatchi Sundaram Muthu Selva Annamalai (University College London), Emiliano De Cristofaro (University of California)
CodeSafty and PrivacyConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Proposes an approximate tight auditing method for DP-SGD under a black-box threat model, significantly amplifying privacy leakage using 'worst-case' initial model parameters;
Bhavya Sukhija (ETH Zurich), Andreas Krause (ETH Zurich)
CodeOptimizationReinforcement Learning
🎯 What it does: A model-based episodic exploration algorithm NEORL is proposed for episodic RL in continuous nonlinear dynamics, optimizing average cost and achieving robust learning.
Neuc-MDS: Non-Euclidean Multidimensional Scaling Through Bilinear Forms
Chengyuan Deng (Rutgers University), Cheng Xin (Rutgers University)
CodeImageTabular
🎯 What it does: Developed Neuc-MDS, which provides a method to extend traditional Multidimensional Scaling (MDS) to non-Euclidean, non-metric similarity using symmetric bilinear forms and negative eigenvalues for low-dimensional embedding.
Justin Dumouchelle (University of Toronto), Elias Boutros Khalil
CodeOptimizationGraph Neural NetworkTabular
🎯 What it does: A dual-layer optimization framework NEUR2BILO based on neural networks is proposed, which transforms the original mixed-integer nonlinear bilevel problem into a single-layer problem that can be solved using MIP, providing a fast heuristic solution method.
Neural Characteristic Activation Analysis and Geometric Parameterization for ReLU Networks
Wenlin Chen (University of Cambridge), Hong Ge (University of Cambridge)
CodeOptimizationImageTabular
🎯 What it does: This paper proposes and analyzes the feature activation boundary of ReLU networks, discovering that common parameterization/normalization can lead to instability in stochastic optimization, and introduces a geometric parameterization GmP in spherical coordinates.
🎯 What it does: This paper proposes a neural network-based combinatorial optimization framework for solving robust path planning with uncertain travel times (RTSP and RVCRP). It extracts features from upper and lower bound information through a dual-head cross-attention mechanism and employs a reinforcement learning (REINFORCE + POMO) training strategy to quickly compute maximum returns using a pre-trained TSP model.
🎯 What it does: A framework is proposed that utilizes pre-trained generative models (DDIM or GAN) to perform gradient optimization in the latent space, thereby generating cover images that are most suitable for hidden information;
Neural Embeddings Rank: Aligning 3D latent dynamics with movements
Chenggang Chen (Johns Hopkins University), Xiaoqin Wang (Johns Hopkins University)
CodeExplainability and InterpretabilityRepresentation LearningContrastive LearningTime Series
🎯 What it does: This paper proposes the Neural Embeddings Rank (NER) method, which embeds neural dynamics into a three-dimensional latent space and aligns continuous motion labels through ranking.
Neural Krylov Iteration for Accelerating Linear System Solving
Jian Luo (MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition University of Science and Technology of China), Yufei Kuang (MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition University of Science and Technology of China)
CodeOptimizationComputational Efficiency
🎯 What it does: The NeurKItt method is proposed, which uses neural operators to predict the invariant subspace of linear systems and utilizes this subspace to accelerate Krylov subspace iterations, significantly reducing the number of iterations and computation time.
Mirco Giacobbe (University of Birmingham), Michael Tautschnig (Amazon Web Services)
CodeOptimizationComputational EfficiencyTabular
🎯 What it does: A model checking method based on neural networks is proposed, utilizing the quantized neural ranking function obtained from training as a formal proof certificate to formally verify the linear temporal logic (LTL) properties of hardware designs.
🎯 What it does: A reparameterization method based on neural networks is proposed to replace traditional coarse-graining, which can both reduce and increase degrees of freedom, directly minimizing energy in the CG space.
Neural P$^3$M: A Long-Range Interaction Modeling Enhancer for Geometric GNNs
Yusong Wang (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)
CodeGraph Neural NetworkGraph
🎯 What it does: The Neural P3M framework is proposed to introduce grid points in geometric GNNs to model long-range interactions and facilitate information exchange between atoms and grids.
🎯 What it does: A framework for learning latent dynamics of point cloud topological features using time evolution (Neural Persistence Dynamics) is proposed, and it is used to infer the parameters of collective behavior models.
🎯 What it does: A learnable gated residual mechanism (Neural-RDM) is proposed, unifying manifolds and U-shaped residual networks, replacing the manually designed mean-variance scheduler to achieve deep scalable diffusion models.
🎯 What it does: This paper presents NeuralFuse, a pluggable input transformation module that can restore the inference accuracy of deep networks when low voltage causes random bit flips in SRAM.
NeuralSolver: Learning Algorithms For Consistent and Efficient Extrapolation Across General Tasks
Bernardo Esteves (Instituto Superior Tecnico, Universidade de Lisboa), Francisco S. Melo (Instituto Superior Tecnico, Universidade de Lisboa)
CodeRecurrent Neural NetworkSequential
🎯 What it does: We propose NeuralSolver, a recursive solver that achieves efficient and robust extrapolation on tasks of both the same and different sizes.
CodeSafty and PrivacyComputational EfficiencyTransformerTextBenchmark
🎯 What it does: A two-party secure inference framework named Nimbus is proposed for Transformer models, providing efficient implementations for both linear and nonlinear layers.
No Free Lunch in LLM Watermarking: Trade-offs in Watermarking Design Choices
Qi Pang (Carnegie Mellon University), Virginia Smith (Carnegie Mellon University)
CodeTransformerLarge Language ModelText
🎯 What it does: This paper studies and demonstrates three types of attacks (de-watermarking, forgery, API attacks) against watermarking for large language models (LLMs) and their impacts on robustness, usability, and security, proposing corresponding defense measures and design guidelines.
No Regrets: Investigating and Improving Regret Approximations for Curriculum Discovery
Alexander Rutherford (University of Oxford), Jakob Nicolaus Foerster (University of Oxford)
CodeRobotic IntelligenceReinforcement Learning
🎯 What it does: This paper studies and improves the task prioritization method in Unsupervised Environment Design (UED), proposing to select training environments directly based on the 'learnability' of tasks to enhance the robustness of RL agents.
No Representation, No Trust: Connecting Representation, Collapse, and Trust Issues in PPO
Skander Moalla (Ecole Polytechnique Federale de Lausanne), Caglar Gulcehre (Ecole Polytechnique Federale de Lausanne)
CodeOptimizationReinforcement LearningSequential
🎯 What it does: This paper studies the relationship between feature representation degradation and performance collapse of PPO in non-stationary training environments, and proposes a Proximal Feature Optimization (PFO) auxiliary loss based on feature pre-activation to alleviate this issue.
🎯 What it does: The FUNGI method is proposed, which combines the output embeddings of pre-trained models with self-supervised gradients to improve k-nearest neighbor retrieval.
Noise-Aware Differentially Private Regression via Meta-Learning
Ossi Räisä (University of Helsinki), Richard E. Turner (University of Cambridge)
CodeSafty and PrivacyMeta LearningConvolutional Neural NetworkTabular
🎯 What it does: This paper proposes a differential privacy convolutional conditional neural process (DPConvCNP) based on meta-learning, which incorporates a functional DP mechanism during meta-training on non-sensitive simulated data. Ultimately, it achieves calibrated and accurate predictions with a single forward inference on real private data while satisfying DP privacy guarantees.
NoiseGPT: Label Noise Detection and Rectification through Probability Curvature
Haoyu Wang (Beijing Institute of Technology), Tongliang Liu (University of Sydney)
CodeClassificationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringImageMultimodality
🎯 What it does: Utilizing a multimodal large language model (MLLM) to detect and correct label noise in image datasets through probability curvature differences, a zero-shot noise detection and correction framework based on In-Context Discrepancy (ICD) is proposed.
🎯 What it does: This paper proposes a noise label learning method based on multiple annotators and sparse constraints, COINNet, which can identify and correct noisy labels in the presence of instance-related noise (considered as outliers), achieving identifiable learning for noise-free classifiers.
🎯 What it does: This paper conducts a non-asymptotic convergence analysis of stochastic gradient descent (SGD) with biased gradients and adaptive step sizes, providing a linear convergence rate under the PL condition and an O(log n / √n + b_n) convergence rate in general non-convex smooth scenarios. It proves that Adagrad, RMSProp, and AMSGRAD can still converge to critical points in the presence of biased gradients.
🎯 What it does: This paper proposes a non-asymptotic uncertainty quantification method based on debiasing estimation and data-driven correction, providing confidence intervals with provable coverage under finite samples.
Yuanqing Wang (New York University), Kyunghyun Cho (Genetech)
CodeRepresentation LearningDrug DiscoveryRecurrent Neural NetworkGraph Neural NetworkGraphBiomedical Data
🎯 What it does: A convolution-free graph neural network RUM is proposed, which constructs node representations by integrating semantic and topological features through random walks and RNN.
Non-geodesically-convex optimization in the Wasserstein space
Hoang Phuc Hau Luu (University of Helsinki), Arto Klami (University of Helsinki)
CodeOptimization
🎯 What it does: A 'semi-forward-backward Euler' discretization method is proposed and analyzed for solving non-geodesically convex non-convex optimization problems with a differential convex (DC) form objective function in Wasserstein space;
Nonlinear dynamics of localization in neural receptive fields
Leon Lufkin (Yale University), Erin Grant (Gatsby Unit and Sainsbury Wellcome Centre University College London)
CodeImage
🎯 What it does: The study investigates the theoretical dynamics of single neuron and multi-neuron models by learning to produce localized receptive fields in feedforward networks without explicit sparse constraints, and validates this through simulation experiments.
Nonlocal Attention Operator: Materializing Hidden Knowledge Towards Interpretable Physics Discovery
Yue Yu (Lehigh University), Stewart A Silling
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyAuto EncoderTabularPhysics Related
🎯 What it does: A Nonlocal Attention Operator (NAO) based on attention is proposed, achieving unified learning for forward prediction and inverse mechanism discovery of physical systems.
Nonparametric Instrumental Variable Regression through Stochastic Approximate Gradients
Yuri Fonseca, Yuri Saporito
CodeOptimizationTabular
🎯 What it does: A non-parametric instrumental variable regression algorithm based on Stochastic Approximate Gradient Descent (SAGD-IV) is proposed, which directly minimizes the overall risk;
Zicheng Sun (Renmin University of China), Feng Zhou (Renmin University of China)
CodePoint CloudTime Series
🎯 What it does: This paper proposes a Non-Stationary Sparse Spectral Poisson Process (NSSPP) and its deep variant (DNSSPP) for flexible modeling of point process intensity functions.
🎯 What it does: This paper proposes the Composition-Incremental Learning (Composition-IL) task and designs the CompILer model to achieve incremental learning of state-object combinations.
🎯 What it does: This study investigates the zero-shot motion segmentation task, comparing the performance of traditional optical flow models with neuroscience-inspired motion energy models on random point stimuli.
Observational Scaling Laws and the Predictability of Langauge Model Performance
Yangjun Ruan (University of Toronto), Tatsunori Hashimoto (Stanford University)
CodeComputational EfficiencyLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes an Observational Scaling Law based on publicly available model standard benchmark scores, allowing for scale predictions across model families without the need to retrain the models.
ODGS: 3D Scene Reconstruction from Omnidirectional Images with 3D Gaussian Splattings
Suyoung Lee (Seoul National University), Kyoung Mu Lee (Seoul National University)
CodeRestorationGenerationComputational EfficiencyGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: A 3D high-quality scene reconstruction method for panoramic images, ODGS, has been developed, utilizing 3D Gaussian splatting for fast rendering.
🎯 What it does: The DARC policy trained using modified rewards in the source domain is transferred to the target domain through observational imitation learning (GAIfO), forming an offline dynamic transfer reinforcement learning framework (DARAIL).
🎯 What it does: This study investigates how to compress massive optimal offline RL data into a small amount of expert behavior data, thereby quickly obtaining high-performance policies through behavior cloning.
Offline Reinforcement Learning with OOD State Correction and OOD Action Suppression
Yixiu Mao (Tsinghua University), Xiangyang Ji (Tsinghua University)
CodeReinforcement LearningTabular
🎯 What it does: This paper proposes an offline reinforcement learning method named SCAS, which unifies the solutions to the problems of out-of-distribution (OOD) states and out-of-distribution (OOD) actions that arise in offline datasets.
🎯 What it does: This paper proposes OmniTokenizer, a transformer-based tokenizer that can uniformly handle images and videos, and achieves visual generation based on this.
CodeClassificationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes and theorizes a framework for affine homotopy comparison between language encoders, defining a semi-metric space and a pre-ordering relation to measure the intrinsic similarity of encoders, and proves that this intrinsic similarity can upper bound the extrinsic similarity of downstream tasks.
On Causal Discovery in the Presence of Deterministic Relations
Loka Li (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Carnegie Mellon University)
CodeDrug DiscoveryScore-based ModelTabular
🎯 What it does: This paper proposes a causal discovery framework called DGES that can address errors caused by violations of faithfulness in traditional methods by detecting deterministic clusters, using an improved Greedy Equivalent Search, and performing local exact searches.
🎯 What it does: This study investigates whether a 'Mesa-optimizer' appears in autoregressive pre-trained transformers and conducts a rigorous analysis of its training dynamics through gradient flow on a single-layer linear self-attention model; it proves that under specific data distribution conditions, the model converges to a structure that implements a single gradient descent and explores its capacity limits.
On provable privacy vulnerabilities of graph representations
Ruofan Wu (Ant Group), Weiqiang Wang (Ant Group)
CodeSafty and PrivacyRepresentation LearningAdversarial AttackGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper studies privacy vulnerabilities in graph representation learning, conducting theoretical analysis and experimental validation on similarity-based edge reconstruction attacks (SERA), and exploring the feasibility of using noise aggregation (NAG) to counter this attack.
🎯 What it does: This paper proposes the Softmax-DPO (S-DPO) loss, which aligns preferences for language models using multiple negative samples to enhance the personalized ranking performance of recommendation systems.
Yongchun Li (University of Tennessee), Weijun Xie (Georgia Tech)
CodeOptimizationTabularBiomedical Data
🎯 What it does: This paper studies the theory and algorithms of Sparse Canonical Correlation Analysis (SCCA), proposing a combinatorial and MISDP formulation and providing both approximate and exact solution methods.
On the Benefits of Public Representations for Private Transfer Learning under Distribution Shift
Pratiksha Thaker (Carnegie Mellon University), Virginia Smith (Carnegie Mellon University)
CodeDomain AdaptationFederated LearningSafty and PrivacyTransformerContrastive LearningImageBiomedical Data
🎯 What it does: The study investigates how public pre-training can enhance the effectiveness of differential privacy (DP) transfer learning in the presence of significant distribution shifts between public and private tasks, and demonstrates its validity through experiments and theoretical proofs.
On the Computational Complexity of Private High-dimensional Model Selection
Saptarshi Roy (University of Michigan), Ambuj Tewari (University of Michigan)
CodeOptimizationSafty and PrivacyTabular
🎯 What it does: Proposed different private optimal subset selection algorithms under high-dimensional sparse linear regression models, implementing model selection by combining the exponential mechanism and Metropolis–Hastings sampling.
On the Expressivity and Sample Complexity of Node-Individualized Graph Neural Networks
Paolo Pellizzoni (Max Planck Institute of Biochemistry), Karsten Borgwardt (Max Planck Institute of Biochemistry)
CodeGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This study analyzes the expressiveness and sample complexity of Graph Neural Networks (GNN) using node personalization techniques, proposing upper bounds on VC dimension and covering number, and designing a low VC dimension EGONN architecture.
On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks
Jiong Zhu (University of Michigan), Danai Koutra (University of Michigan)
CodeGraph Neural NetworkGraph
🎯 What it does: This study investigates the impact of feature heterogeneity on graph neural networks in link prediction and systematically analyzes the design of encoders and decoders at different levels of feature homogeneity.
On the Limitations of Fractal Dimension as a Measure of Generalization
Charlie Tan, Anthea Monod (Imperial College London)
CodeConvolutional Neural NetworkImageTabular
🎯 What it does: Conduct large-scale experimental evaluations of the generalization metric based on persistent homology (PH) dimensions, examining its correlation with generalization error and uncovering potential failure modes.
On the Noise Robustness of In-Context Learning for Text Generation
Hongfu Gao (Southern University of Science and Technology), Hongxin Wei (Southern University of Science and Technology)
CodeGenerationText
🎯 What it does: Analyzes and addresses the decline in contextual learning performance caused by noise in demonstration data for text generation tasks, proposing the Local Perplexity Ranking (LPR) method to enhance noise robustness.
On the Parameter Identifiability of Partially Observed Linear Causal Models
Xinshuai Dong (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
CodeOptimization
🎯 What it does: This paper studies the parameter identifiability of partially observable linear causal models and proposes a new parameter estimation method based on this.
🎯 What it does: A method is proposed and validated that replaces traditional Fine-Tuning by only transferring attention patterns (Attention Transfer), where the student model uses the attention information from the pre-trained teacher while learning its own features, significantly improving the performance of ViT on downstream tasks.
On the Worst Prompt Performance of Large Language Models
Bowen Cao (Chinese University of Hong Kong), Wai Lam (Chinese University of Hong Kong)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a new benchmark, ROBUSTALPACAEVAL, to evaluate the worst performance of LLMs under prompts with varying semantic equivalence and fluency, and conducts large-scale experiments on ChatGPT and six open-source LLMs.
One Token to Seg Them All: Language Instructed Reasoning Segmentation in Videos
Zechen Bai (National University of Singapore), Mike Zheng Shou (National University of Singapore)
CodeSegmentationTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: This paper presents VideoLISA, a video instruction reasoning segmentation model based on a multimodal large language model, capable of generating cross-frame semantic segmentation masks based on language descriptions.
One-Shot Safety Alignment for Large Language Models via Optimal Dualization
Xinmeng Huang (University of Pennsylvania), Dongsheng Ding (University of Pennsylvania)
CodeOptimizationSafty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A one-shot safe alignment method is proposed, transforming the constrained language model alignment problem into an unconstrained alignment problem. The optimal Lagrange multipliers are solved through a closed-form dual function, achieving compliance with safety constraints with only one training session.
🎯 What it does: This paper proposes a data-free distillation method called Score Implicit Matching (SIM), which compresses multi-step diffusion models into a single-step generator.
🎯 What it does: OSEDiff is proposed, a one-shot diffusion network that uses low-quality images directly as the diffusion starting point to recover high-quality images, eliminating the uncertainty of random noise and significantly reducing computational costs.
OneBit: Towards Extremely Low-bit Large Language Models
Yuzhuang Xu (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
CodeCompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a OneBit framework that compresses LLM weights to 1-bit and maintains accuracy through bidirectional vectors, achieving extremely low-bit quantization.
OneRef: Unified One-tower Expression Grounding and Segmentation with Mask Referring Modeling
Linhui Xiao (Institute of Automation Chinese Academy of Sciences), Changsheng Xu (Institute of Automation Chinese Academy of Sciences)
CodeRecognitionObject DetectionSegmentationTransformerVision Language ModelImageMultimodality
🎯 What it does: Proposes the OneRef framework and Mask Referring Modeling (MRefM) technique, utilizing a single-tower shared Transformer for unified encoding of vision and text, achieving direct regression for expression localization and segmentation.
Online Adaptation of Language Models with a Memory of Amortized Contexts
Jihoon Tack (KAIST), Jonathan Richard Schwarz (Harvard University)
CodeDomain AdaptationOptimizationMeta LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an online adaptation framework based on Memory-Augmented Context (MAC), which can quickly adapt large language models (LLMs) by compressing new documents into learnable modulations (PEFT) and storing them in a memory bank, allowing for the aggregation of a single modulation during query time without updating the large model parameters.
🎯 What it does: An online label shift adaptation framework OLS-OFU is proposed, which enhances prediction performance in online label shift and generalized label shift scenarios by dynamically updating the feature extractor using self-supervised learning (SSL) during the testing phase.
🎯 What it does: An online relational inference framework ORI is proposed, which utilizes a trainable adjacency matrix and online backpropagation to identify the hidden interaction graph of multi-agent systems in real-time.
🎯 What it does: An online temporal action segmentation framework is proposed, which includes an adaptive memory bank, a context-aware feature enhancement module, and post-processing methods for action segmentation of untrimmed videos without a look-back window.
Only Strict Saddles in the Energy Landscape of Predictive Coding Networks?
Francesco Innocenti (University of Sussex), Christopher Buckley
CodeOptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the energy landscape of the Predictive Coding Network (PCN) at the inference equilibrium point, derives a closed-form expression for deep linear networks, and proves that many originally non-strict saddle points become strict saddle points under equilibrium energy. Experiments confirm the theoretical predictions and propose the conjecture that all saddle points are strict saddle points.
Hefei Li (East China Normal University), Zhengfeng Yang (East China Normal University)
CodeGraph Neural NetworkBenchmark
🎯 What it does: An open-book neural algorithm inference framework is proposed, allowing the network to access and utilize instances from the training set during the inference process.
🎯 What it does: The OptEx framework is proposed, which utilizes kernelized gradient estimation based on historical gradients to achieve approximate parallel iterations of first-order optimization (FOO) algorithms, significantly reducing the required number of sequential iterations.
🎯 What it does: This paper proposes modeling the performative effect as a push-forward operation and provides a corresponding gradient estimation method; it proves that linear shifts can lead to the convexity of performative risk in the binary classification case and relates it to robust classification; it also introduces the RPPerfGD algorithm based on gradient descent, which enhances performance by learning the shift matrix.
🎯 What it does: Developed Optimal Flow Matching (OFM), which learns the optimal transport flow of straight-line trajectories in single-step Flow Matching, allowing for direct acquisition of straight paths without solving ODEs;
Optimistic Critic Reconstruction and Constrained Fine-Tuning for General Offline-to-Online RL
Qin-Wen Luo (Nanjing University of Aeronautics and Astronautics), Sheng-Jun Huang (Nanjing University of Aeronautics and Astronautics)
CodeReinforcement LearningSequential
🎯 What it does: A general offline-to-online reinforcement learning framework is designed, which first reconstructs the critic through optimistic re-evaluation and aligns it with the offline policy, then incorporates constraint fine-tuning to alleviate distribution shift, achieving stable and efficient online fine-tuning from any offline method to SAC, TD3, and PPO.
Optimistic Verifiable Training by Controlling Hardware Nondeterminism
Megha Srivastava (Stanford University), Dan Boneh (Stanford University)
CodeTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes a verifiable training method that eliminates non-determinism between different GPUs by training at higher precision and recording rounding decisions, allowing auditors to accurately reproduce the training process.
Optimization Algorithm Design via Electric Circuits
Stephen P. Boyd (Stanford University), Jaewook J. Suh (Rice University)
CodeOptimizationTabular
🎯 What it does: This paper proposes a convex optimization algorithm design method based on circuit simulation, mapping the continuous-time dynamics of optimization problems to an inductor-capacitor-resistor (RLC) circuit, and utilizing automatic discretization techniques to generate convergent algorithms.